How to Optimize for Extreme Scale Problems?

Global models work really well for these extreme scale problems. But if you were to extract just a single random set of forecasts from that global model, the discrete ones are going to beat it every single time on accuracy. We have customers at Databricks that are doing this today that are running 4.5 million profit models in production. The real trick is, how do you solve that problem where you need accuracy and you also need to store the model?